[Frontend] Support count_reasoning_tokens in the Streaming Parser Engine (#45802)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Chauncey
2026-08-15 06:51:14 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 5cecfc0137
commit ac2ae8798c
16 changed files with 669 additions and 73 deletions
+19 -1
View File
@@ -47,6 +47,11 @@ class ParserEngineReasoningAdapter(ReasoningParser):
def __init__(self, tokenizer: TokenizerLike, *args, **kwargs) -> None:
super().__init__(tokenizer, *args, **kwargs)
self._parser_engine = self._parser_engine_cls(tokenizer, **kwargs) # type: ignore[call-arg]
self._parser_engine_kwargs = kwargs
self._counting_parser_engine: ParserEngine | None = None
# TODO: Remove once Responses finalization reuses accumulated streaming
# parser results instead of reparsing the complete output.
self._streaming_count_valid = False
@contextmanager
def _skip_tool_parsing(self) -> Iterator[None]:
@@ -71,6 +76,7 @@ class ParserEngineReasoningAdapter(ReasoningParser):
model_output: str,
request: ChatCompletionRequest | ResponsesRequest,
) -> tuple[str | None, str | None]:
self._streaming_count_valid = False
with self._skip_tool_parsing():
return self._parser_engine.extract_reasoning(model_output, request)
@@ -83,6 +89,7 @@ class ParserEngineReasoningAdapter(ReasoningParser):
current_token_ids: Sequence[int],
delta_token_ids: Sequence[int],
) -> DeltaMessage | None:
self._streaming_count_valid = True
with self._skip_tool_parsing():
return self._parser_engine.extract_reasoning_streaming(
previous_text,
@@ -122,7 +129,18 @@ class ParserEngineReasoningAdapter(ReasoningParser):
return self._parser_engine.get_streaming_fallback_content(text, request)
def count_reasoning_tokens(self, token_ids: Sequence[int]) -> int:
return self._parser_engine.count_reasoning_tokens(token_ids)
if self._streaming_count_valid:
return self._parser_engine.count_reasoning_tokens(token_ids)
if not token_ids:
return 0
if self._counting_parser_engine is None:
self._counting_parser_engine = self._parser_engine_cls(
self.model_tokenizer, **self._parser_engine_kwargs
) # type: ignore[call-arg]
self._counting_parser_engine._single_pass_parse(
self.model_tokenizer.decode(token_ids), token_ids
)
return self._counting_parser_engine.count_reasoning_tokens(token_ids)
class ParserEngineToolAdapter(ToolParser):